Refusing to guess

A plan set is ambiguous in places. Two grade lines cross the centreline within a few tenths of a foot and the sheet never says which is design. A station label is clipped at the page edge. A datum cannot be corroborated by any second source. These are not rare pathologies; they are Tuesday.

The tempting behaviour

Pick the more likely one. Interpolate. Take the average. Each is defensible in isolation and catastrophic in aggregate, because the output looks identical to a confident, correct answer. A number with no marker on it will be trusted exactly as much as a number that was actually read.

A guess and a measurement look the same in a spreadsheet. That is the whole problem.

What refusal means here

When the evidence does not determine a value, Mathyra emits a refusal instead: a named reason from a fixed vocabulary, attached to the location it applies to, counted in the denominator of whatever was being measured. Refusals are not errors swept aside — they are results, and they are visible in the conformance report beside everything that did resolve.

• Every refusal reason is enumerated and defined; new ones cannot appear silently
• Refused items keep their place in the count, so coverage is never inflated by dropping the hard cases
• A refusal carries the evidence that triggered it, so a human can adjudicate it in seconds
• Where two readings disagree, the disagreement is reported — not averaged away

The cost, stated plainly

This is expensive. On one reference set, applying a stricter datum rule meant roughly three in ten exported coordinates stopped shipping. That is a real reduction in output, and we recorded it as such rather than quietly loosening the rule.

The trade seems right to us for one reason: an engineer can work with "we could not read this." Nobody can work with a number that was invented and looks fine.